AI implementation for a Wildberries & Ozon seller: 26 automations in 6 weeks

An AI implementation case for a marketplace seller on Wildberries and Ozon: a cosmetics manufacturer with six seller accounts and about ₽45M in monthly turnover. In six weeks a six-person team launched 26 automations instead of the three planned, and the owner now sees the whole company every morning. Below is the interim measurement with its method: what the AI environment delivered and what came from the season, prices and the marketplaces.

Published October 1, 2026 · Breakdown by Anton Ro Baten

Point A: where they started

• ·No consolidated reporting: numbers were pulled together by hand for each specific question.
• ·Sales were checked twice a week, costs once a month.
• ·No single product catalogue — duplicates and mismatches across seller accounts.
• ·Answering “what is this account’s margin for the period” took a day to a week through the finance person.
• ·Every key decision went through the owner, who runs the company remotely.

What we did

1. 1

   A corporate AI environment for eight people

   The owner, the COO, the head of sales, the account managers and an administrator work in one secure environment. Each has a workspace describing their role and processes, so the AI assistant knows what the person does from the very first request.

2. 2

   Connections to the data

   Direct access to the Wildberries and Ozon seller accounts, a marketplace analytics service, the accounting system, the task tracker, email, drive and calendar. Accounting is read-only: AI gathers and analyses, people make the decisions.

3. 3

   Role unpacking

   Every employee went through an interview: what they do, how long it takes, where they make mistakes. The answers became a process map and a list of tasks to automate, sorted by impact.

4. 4

   26 automations built by the team

   Summaries, price and stock control, ad and financial report analysis, review replies, a “price traffic light”, paid storage. Each employee launched their automation themselves and defended it in front of the COO with a screen demo — acceptance inside the company, not by a contractor.

5. 5

   A morning summary for the owner

   Every day at 6:00 — six seller accounts on one screen: revenue, profit, margin, ad share of revenue, loss-making products, campaigns burning budget and stock about to run out.

What AI found in the data

• ·The company’s own brand was listed as a competitor in the market report: its niche share was understated threefold — 12.9% instead of 37.8%. The company is the niche leader, not number three.
• ·No unit growth in a year: 59 thousand units a month then and now. All of the 27% revenue growth is price, and demand did not react to the increase.
• ·One marketplace’s margin halved in a year — from 26% to 14% — and its share of revenue fell from 52% to 26%.
• ·Racks, machines and compressors worth ₽4M are booked as goods for sale, which distorts inventory and turnover figures.
• ·Certificates of conformity covering 43% of revenue expire on the same day, including the two best-selling products.
• ·The margin on identical products differs across accounts — from 13% to 25%.
• ·Logistics in one account costs ₽98.6 per unit versus ₽62.2 for the same products in another: that is 56% of the account’s profit.
• ·The company’s own accounts compete with each other in ad auctions — 10 of its own products in one search result.
• ·Dead stock: an item with 1,128 days of inventory.
• ·An access key to the finances of all six accounts sat in plain text in a document shared by link. The key was revoked.

Point B: results

Daily profit grew by a fifth even though the marketplaces took 11 more percentage points of revenue — about ₽6.5M a month at September volumes. Sales growth is not counted in the payback.

How we measured

• ·The baseline is June: the last month both marketplaces ran without disruptions. Wildberries warehouses were disrupted in the summer, and comparing with July or August would inflate the result.
• ·September is incomplete and buyouts have not matured, so everything is per day; profit may still be adjusted.
• ·Figures come from a marketplace analytics service with one method for all periods. Profit is the marketplace department’s profit after taxes.
• ·Hours are the employees’ own estimates, converted to money at salary rates without bonuses.
• ·Every figure is tagged with its source: from the system, an estimate or a calculation. Where there is no figure, it says “not measured”, not a guess.

What AI does not get credit for

• ·Revenue and sales growth — the season, warehouse deliveries and the marketplace recovering from disruptions.
• ·Average order value growth — pricing decisions were made by people.
• ·The commission increase — the marketplaces’ decision; the environment only measured it.
• ·What the environment actually delivered is speed, frequency and a complete picture: sales and costs every day, a margin answer in minutes instead of days.

What didn’t work

• ·The losses found are not stopped yet: the signal arrives, the action does not. One loss-making campaign kept running after two warnings in the morning summary.
• ·Four of six people hit the AI usage limits; the head of sales lost up to half a working day to it.
• ·One automation was stopped — its numbers did not match the seller account.
• ·Ozon reviews cannot be automated without a separate marketplace plan.

In their own words

What’s next

• ·An execution loop: every finding gets a task, an owner, a deadline and a check in the weekly review that the loss is gone.
• ·Clean inventory accounting: move equipment out of goods for sale.
• ·Prices adjusted to the new commission — starting with the account whose margin halved.
• ·A repeat measurement in three months: what was stopped and what it brought in rubles.

Takeaways for your business

1. 1.In the first weeks AI is an X-ray of your data: the most valuable things it finds are errors in accounting and reports, not automations.
2. 2.The best automations come from the employee whose routine it is. Acceptance belongs to their manager — then the tool stays in the company, not with a contractor.
3. 3.Measure from an honest baseline and don’t credit AI with the season or prices: a figure without a source is not a figure.
4. 4.Finding a loss is half the job: every signal needs an owner and a deadline.